Abstract:Aiming at the limitations of traditional forecasting methods for ultra-short term photovoltaic power, a method based on LightGBM-TextCNN-XGBoost algorithm model for ultra-short term photovoltaic power forecasting is proposed. First, the original data is preprocessed, and the data is modal decomposed through CEEMDAN. After the modal decomposed data is normalized, the data is clustered into three weather types based on the GWO-FCM clustering algorithm. Secondly, the data is divided and the training set trains the LightGBM and TextCNN algorithms respectively. A model based on the LightGBM-TextCNN-XGBoost algorithm is established based on the Stacking idea to predict. The test set is used to evaluate the model prediction effect. Finally, combined with R Squared and other evaluation indexes, the prediction model is comprehensively evaluated. The results showed that the prediction effect was good (RSquared=0.9766). The model proposed in this paper can accurately predict the efficiency of photovoltaic power generation, help photovoltaic power plants reduce losses, and thus ensure the safe and robust operation of microgrids